UVR VR denoise and de-echo models, GGUF

No license is stated for these weights anywhere upstream. They are distributed with Ultimate Vocal Remover's model downloads and mirrored without one. They are republished here in GGUF form with that caveat. Do not assume any license, and if you hold the rights and want them removed, open a discussion on this repository.

Five of Ultimate Vocal Remover's VR-architecture (UVR 5.1 CascadedNet) models, converted to GGUF for stems.cpp, a C++/ggml stem separator that runs on CPU, CUDA, Vulkan and Metal with no Python. Each returns two stems: the one the network's mask selects (primary) and its complement.

F32 only. SHA256SUMS lists every file.

Download and run

git clone --recursive https://github.com/betweentwomidnights/stems.cpp && cd stems.cpp
./build.sh vulkan                          # or cpu | cuda | metal   (Windows: build.cmd)
./models.sh uvr_denoise uvr_deecho_normal
build-vulkan/bin/stems-split -m models/uvr_denoise-32M-v1.0-F32.gguf -i recording.wav -o clean/ --stems denoised
build-vulkan/bin/stems-split -m models/uvr_deecho_normal-32M-v1.0-F32.gguf -i vocal.wav -o dry/ --stems no_echo

stems-server finds each by model id ("model": "uvr_denoise"). The stems.cpp release packages (v0.1.3 and later) load them through the same C ABI as the other models.

What the runtime does

The network is UVR's 5.1 CascadedNet, unchanged. The spectral path around it is defined explicitly so every platform computes the same thing: periodic Hann STFT with centred zero padding, UVR's band layout (4band_v3 for the 32M/56M models, 1band_sr44100_hl1024 for Lite), polyphase resampling between bands with the filters stored in the GGUF, UVR's prefilter and synthesis masks, and 256-frame windows with 64 frames discarded at each edge. It is not byte-identical to the UVR GUI: UVR's Windows synthesis normally uses sinc_fastest, and its optional aggressiveness, test-time augmentation, artifact merging and high-end mirroring are not implemented. See stems.cpp's docs/VR.md.

Parity

SNR of the C++ output against the unchanged UVR network and stems.cpp's portable Python pipeline, on a 4 s stereo music excerpt, measured 2026-10-07/09 with ggml 9d0d910b. CPU: Core Ultra 9 275HX. Vulkan and CUDA: RTX 5070 Laptop. Every run passes stems.cpp's check (cosine >= 0.99999, SNR >= 50 dB).

DeEcho-DeReverb on CUDA is the one outlier, still far below audibility.

On Apple Metal (M4, 32 GB, references dumped there) every model passes too, with cosine 1.000000000. Mask / primary / secondary SNR: De-Echo Normal 118.8 / 122.5 / 110.0 dB, De-Echo Aggressive 116.0 / 121.8 / 112.7 dB, DeEcho-DeReverb 121.9 / 120.3 / 105.8 dB (no sign of the CUDA gap), each in under 1.8 s for the 4 s excerpt. DeNoise passed Metal parity when ggml's im2col change was validated.

SNR here is against reference outputs, not against clean recordings: it shows the port computes what UVR's network computes, not how well a model denoises.

Sources

Checkpoints from the seanghay/uvr_models mirror at revision 6f4fc0c, SHA256-verified by the converter; UVR modelparams and network code from Anjok07/ultimatevocalremovergui at a5f88453bfb2b38b05a965bcf67727243e0cbf19.

Converted with tools/convert_vr.py --preset <denoise_lite|denoise|deecho_normal|deecho_aggressive|deecho_dereverb>. BatchNorm is folded into the convolution and dense weights, the training-only auxiliary head is dropped, and the band layout and resampling filters are written as metadata. The architecture and weights are otherwise unchanged. Conversion is deterministic: re-running it reproduces these files byte for byte.